Our method can be applied to a broader range of tasks that fall into visual

Neural Information Processing Systems 

In [7], only the loss of the last step is used as opposed to the complete trajectory in our method.